Mothers, adults, children, pets - Towards the acoustics of intimacy

In this paper, we investigate acoustic features which differentiate the two speech registers neutral and intimate within different constellations of speakers and addressees. Three different types of speakers are considered: mothers addressing their own children or an unknown adult, women with no children addressing an imaginary child or an imaginary adult, and children addressing a pet robot using both intimate and neutral speech. We use a large, systematically generated feature vector, upsampling, and SVM and RF for learning. Results are reported for extensive test-runs facing speaker-independency and using PCA-SFFS vs. SVM-SFFS for feature ranking. Classification performance and most relevant feature types are discussed in detail.